Do the stimuli of an SSVEP-based BCI really have to be the same as the stimuli used for training it ?

Does the stimulation used during the training on an SSVEP-based BCI have to be similar from the one of the end use ? We recorded 6-channels EEG data from 12 subjects in various conditions of distance between targets, and of difference in color between targets. Our analysis revealed that the stimulation configuration used for training which leads to the best classification accuracy is not always the one which is closest to the end use configuration. We found that the distance between targets during training is of little influence if the end use targets are close to each other, but that training at far distance can lead to a better accuracy for far distance end use (p<0.01). Additionally, an interaction effect is observed between training and testing color (p<0.001) : while training with monochrome targets leads to good performance only when the test context involves monochrome targets as well, a classifier trained on colored targets can be efficient both for colored and monochrome targets. In a nutshell, in the context of SSVEP-based BCI, training using distant targets of different colors seems to lead to the best and more robust performance in all end use contexts.